Adaptive Now‐ and Forecasting of Global Temperatures Under Smooth Structural Changes

Accurate short‐term now‐ and forecasting of global temperatures is an important issue and helpful for policy design and decision making in the public and private sectors. We compose a raw mixed‐frequency data set from weather stations around the globe (1920–2020). First, we document smooth variation in average monthly and annual temperature series by applying a dynamic stochastic coefficient model. Second, we use adaptive cross‐validated forecasting methods, which are robust to smooth changes of unknown form in the short run. Therein, recent and past observations are weighted in a mean‐squared error‐optimal way. Overall, it turns out exponential smoothing methods (with bootstrap aggregation) often perform best. Third, by exploiting monthly data, we propose a simple procedure to update annual nowcasts during a running calendar year and demonstrate its usefulness. In this context, we also discuss the role of forecast reconciliation. Further, we show that these findings are robust with respect to climate zones. Finally, we investigate now‐ and forecasting of climate volatility via a range‐based measure and a quantile‐based climate risk measure.